cs.LGMar 25, 2026

Marchuk: Efficient Global Weather Forecasting from Mid-Range to Sub-Seasonal Scales via Flow Matching

Authors: Arsen Kuzhamuratov, Mikhail Zhirnov, Andrey Kuznetsov, Ivan Oseledets, Konstantin Sobolev

Organizations: AXXX · FusionBrain Lab

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Accurate subseasonal weather forecasting remains a major challenge due to the inherently chaotic nature of the atmosphere, which limits the predictive skill of conventional models beyond the mid-range horizon (approximately 15 days). In this work, we present Marchuk, a generative latent flow-matching model for global weather forecasting spanning mid-range to subseasonal timescales, with prediction horizons of up to 30 days. Marchuk conditions on current-day weather maps and autoregressively predicts subsequent days' weather maps within the learned latent space. We replace rotary positional encodings (RoPE) with trainable positional embeddings and extend the temporal context window, which together enhance the model's ability to represent and propagate long-range temporal dependencies during latent forecasting. Marchuk offers two key advantages: high computational efficiency and strong predictive performance. Despite its compact architecture of only 276 million parameters, the model achieves performance comparable to LaDCast, a substantially larger model with 1.6 billion parameters, while operating at significantly higher inference speeds. We open-source our inference code and model at: https://v-gen-ai.github.io/Marchuk/

Figures & tables

Figure 1: Marchuk operates in the latent space learned by the DC-AE model introduced in LaDCast. Within this latent space, we train a flow-matching Diffusion Transformer (DiT) to model the conditional distribution of future weather fields given the current state, enabling efficient and accurate probabilistic forecasting.
Figure 2: Marchuk is conditioned on weather maps from the previous KK days and is provided with noised weather maps for the subsequent NN days as input. The model generates refined forecasts for the next NN days, effectively denoising and predicting future weather states.
Figure 3: Marchuk is flow matching DiT model with additional architecture optimizations.
Model# params#diffusion stepsSpeed, min
LaDCast375M1011
LaDCast1.6B1045
Marchuk276M207.5
Table 1: Inference Speed. The model performs predictions entirely in the latent space for a 30-day forecast horizon with an ensemble size of 50, executed on an H100 GPU. Reported measurements account only for the forecasting component, excluding encoding / decoding.
MethodUW500T500G500SH500 , 10−310^{-3}SLP10m-UW10m-VWT2M
LaDCast, 375M8.143.07784.210.832690.743.813.932.59
LaDCast, 1.6B8.083.04774.180.796684.293.773.912.54
Marchuk, 276M8.153.08785.910.803697.893.803.952.52
Climatology8.553.18810.570.857711.213.964.022.66
Table 2: Comparison to LaDCast. We report RMSE metrics at 15-day forecast horizons. The evaluated variables include atmospheric fields – UW500 (u-component of wind at 500 hPa), T500 (temperature at 500 hPa), G500 (geopotential at 500 hPa), and SH500 (specific humidity at 500 hPa) – as well as surface fields – SLP (sea level pressure), 10m-UW and 10m-VW (u- and v-components of wind at 10 meters), and T2M (temperature at 2 meters).
MethodUW500T500G500SH500 , 10−310^{-3}SLP10m-UW10m-VWT2M
LaDCast, 375M8.473.26840.580.883727.403.934.012.80
LaDCast, 1.6B8.403.17815.530.825713.083.873.972.66
Marchuk, 276M8.403.16809.150.836715.793.863.992.64
Climatology8.553.18810.570.857711.213.964.022.66
Table 3: Comparison to LaDCast. We report RMSE metrics at 30-day forecast horizons. The evaluated variables include atmospheric fields – UW500 (u-component of wind at 500 hPa), T500 (temperature at 500 hPa), G500 (geopotential at 500 hPa), and SH500 (specific humidity at 500 hPa) – as well as surface fields – SLP (sea level pressure), 10m-UW and 10m-VW (u- and v-components of wind at 10 meters), and T2M (temperature at 2 meters).
Figure 4: RMSE comparison. We evaluate LaDCast and Marchuk on the WeatherBench-2 benchmark over a 30-day prediction horizon. The atmospheric variables include the u-component of wind, temperature, geopotential, and specific humidity at 500, 750, and 850 hPa levels. Surface variables include mean sea level pressure, 10-meter wind u- and v-components, and temperature at 2 meters. Marchuk outperforms the small LaDCast model and achieves comparable performance to the large LaDCast model. Beyond approximately 30 days, both models’ forecasts converge toward climatology, indicating that extending the prediction horizon further provides limited practical value.
Figure 5: CRPS Ensemble Metrics Comparison. Figure illustrates the evolution of CRPS over a 30-day forecast horizon. The first row presents results for atmospheric variables at the 500 hPa level, including the u-component of wind, temperature, geopotential, and specific humidity. The second row shows CRPS for surface variables, namely mean sea level pressure, the 10-meter u-component of wind, 2-meter temperature, and total accumulated precipitation over 6 hours.
MethodUW500T500G500SH500 , 10−310^{-3}SLP10m-UWT2MTP-6h, 10−310^{-3}
LaDCast, 375M4.231.49354.070.394305.531.851.130.520
LaDCast, 1.6B4.191.45338.430.377299.991.821.090.515
Marchuk, 276M4.261.47343.340.385310.331.851.100.519
Table 4: Comparison to LaDCast. We report CRPS metrics at 15-day forecast horizons. The evaluated variables include atmospheric fields – UW500 (u-component of wind at 500 hPa), T500 (temperature at 500 hPa), G500 (geopotential at 500 hPa), and SH500 (specific humidity at 500 hPa) – and surface fields – SLP (sea level pressure), 10m-UW (u-component of wind at 10 meters), T2M (temperature at 2 meters), and TP-6h (total accumulated precipitation over the last 6 hours).
MethodUW500T500G500SH500 , 10−310^{-3}SLP10m-UWT2MTP-6h, 10−310^{-3}
LaDCast, 375M4.431.64398.000.419325.581.921.240.530
LaDCast, 1.6B4.381.53364.100.393316.721.881.150.522
Marchuk, 276M4.421.53354.930.401318.991.891.160.527
Table 5: Comparison to LaDCast. We report CRPS metrics at 30-day forecast horizons. The evaluated variables include atmospheric fields – UW500 (u-component of wind at 500 hPa), T500 (temperature at 500 hPa), G500 (geopotential at 500 hPa), and SH500 (specific humidity at 500 hPa) – and surface fields – SLP (sea level pressure), 10m-UW (u-component of wind at 10 meters), T2M (temperature at 2 meters), and TP-6h (total accumulated precipitation over the last 6 hours).
Figure 6: Extreme-Event Forecasting Ability. We evaluate the model’s capability to predict extreme events by analyzing the probability of a cold-weather event preceding the Moscow frost day (midnight, January 26, 2026). Specifically, we plot the probability as a function of lead time, defined as the number of days before the event. The probability is computed as the fraction of ensemble members predicting temperatures within the range of −22∘C±3∘C-22^{\circ}C\pm 3^{\circ}C . This analysis provides insight into the model’s ability to capture the likelihood of extreme temperature events over extended forecast horizons.
ConfigUW500T500G500SH500 , 10−310^{-3}SLP10m-UW10m-VWT2M
RoPE 3D8.633.23829.830.851730.813.984.112.66
GeoRoPE8.533.20821.280.848729.823.974.092.67
RoPE 1D + trainable 2D8.423.18808.110.841719.793.944.052.65
Table 6: Evaluation of Marchuk with different positional encoding schemes: RoPE3D, GeoRoPE, and RoPE 1D with trainable spatial embeddings. The hybrid configuration achieves the best performance in terms of RMSE over a 15-day forecast horizon.
ConfigUW500T500G500SH500 , 10−310^{-3}SLP10m-UW10m-VWT2M
6 hours8.613.21820.050.864727.834.024.102.68
12 hours8.543.15806.220.842714.673.934.082.63
24 hours8.513.17809.980.834715.943.954.082.66
48 hours8.393.14804.050.828713.983.934.082.61
Table 7: Effect of fixed-window training on model performance. Increasing the context window length consistently improves RMSE across all evaluated variables.
ConfigUW500T500G500SH500 , 10−310^{-3}SLP10m-UW10m-VWT2M
1 hour8.603.20829.130.847733.353.984.102.72
6 hours8.493.14814.460.834717.283.944.052.65
12 hours8.653.32851.860.858744.544.024.212.75
Table 8: Effect of temporal skip interval on model performance. Training with different timestep intervals (1, 6, and 12 hours) shows that a 6-hour interval achieves the best RMSE performance, balancing temporal resolution and autoregressive efficiency.
ConfigUW500T500G500SH500 , 10−310^{-3}SLP10m-UW10m-VWT2M
1 day8.433.15808.560.827717.273.924.062.68
2 days8.343.13800.600.816715.863.924.042.59
4 days8.313.15815.620.816728.413.894.082.57
8 days8.533.20826.770.817741.443.934.112.65
Table 9: Impact of inference configuration on model performance. Varying the input context length from 1 to 8 days shows that intermediate horizons (2-4 days) yield the best RMSE performance.
ConfigUW500T500G500SH500 , 10−310^{-3}SLP10m-UW10m-VWT2M
w/o timestamp8.393.14804.050.828713.983.934.082.61
w timestamp8.393.14799.180.827708.663.924.052.59
Table 10: Ablation study on the effect of including timestamp information in Cross-Attention. Incorporating timestamp as a conditioning signal improves RMSE across all evaluated variables.
ConfigUW500T500G500SH500 , 10−310^{-3}SLP10m-UW10m-VWT2M
12 hours8.503.17822.710.838739.364.004.122.62
1 day8.423.18811.020.839727.003.964.092.65
2 days8.513.20817.010.835715.083.964.092.65
Table 11: Ablation study on input context length. Reducing the context to 4 points (1 day of weather maps) maintains predictive accuracy while improving computational efficiency.
Figure 7: ACC score. We evaluate model performance using ACC across several weather variables. The results, illustrated in the figure, show that our model exhibits the smallest decline in ACC over time compared to both LaDCast models.
Figure 8: The CRPS Skill score evaluates how closely ensemble predictions align with the observed outcomes on average. Our model outperforms both the small and large LaDCast variants, demonstrating improved probabilistic forecasting accuracy.
Figure 9: CRPS Spread. The CRPS spread reflects the diversity among ensemble members by measuring pairwise differences within the ensemble. Our model exhibits lower spread compared to LaDCast, indicating reduced ensemble diversity. This effect can be attributed to stronger conditioning: while LaDCast conditions on a single weather map, Marchuk conditions on the four most recent maps, leading to more constrained and less diverse predictions.
More in this topic

Weather Forecasting

Weather forecasting predicts future atmospheric conditions, from local storms to global patterns, using physical models, observations, and machine learning.

See all 176 papers

Explore similar work

CardsList

  1. Timestep-Conditioned Transformers for Global Weather Forecasting

    Aug 6, 2026Sam Levang, Fran Bartolic, Ty Dickinson +3Time Series ForecastingProbabilistic Forecasting

    Salient

  2. AIFS-SUBS: Extending Data-Driven Forecasting to Sub-Seasonal Timescales

    Jul 6, 2026Jakob Schloer, Steffen Tietsche, Christopher D. Roberts +6Probabilistic ForecastingWeather Forecasting

    1European Centre for Medium-Range Weather Forecasts (ECMWF)

  3. S2S-JEPA: Predicting the Predictable at Subseasonal-to-Seasonal Timescales

    Oct 2, 2026Chenyu Dong, Gianmarco MengaldoWeather ForecastingSpatiotemporal Forecasting

    Department of Mechanical Engineering, College of Design and Engineering, National University of Singapore, 9 Engineering Drive 1, 117575, Singapore.